-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtraining.py
132 lines (112 loc) · 6.38 KB
/
training.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
import torch
import os
import tqdm
import wandb
import cv2
import logging
# LUT training pipeline.
def train(model, train_dataloader, loss_fn, optimizer, scheduler, args):
print("Start training LUT model.")
logging.info("Start training LUT model.")
loss_best = 1e10
# Training.
model.train()
acc_loss = 0
tqdm_bar = tqdm.tqdm(total=args.iterations, desc="Training", position=0)
for batch_idx, (global_img, crop_aug, crop_mask, crop_gt, lut_reverse, crop_aug_reverse, is_dual_lut) in enumerate(
train_dataloader):
# Transfer to GPU.
global_img = global_img.cuda()
crop_aug = crop_aug.cuda()
crop_mask = crop_mask.cuda()
crop_gt = crop_gt.cuda()
lut_reverse = lut_reverse.cuda()
crop_aug_reverse = crop_aug_reverse.cuda()
is_dual_lut = is_dual_lut.cuda()
# Forward.
fit_lut3d, lut_transform_image = model(crop_aug, crop_mask)
# Compute loss.
loss_lut_transform_image = loss_fn['masked_mse'](lut_transform_image, crop_gt, crop_mask)
loss_lut_regularize = loss_fn['regularize_LUT'](fit_lut3d)
# Since the LUT inversion value is not perfect (usually color inconsistency), we lower the weight of this loss.
loss_lut_value = loss_fn['mse'](fit_lut3d.permute(0, 2, 3, 4, 1), lut_reverse, is_dual_lut) * 0.1
loss_tv, loss_mn = loss_fn['TV_3D'](args.LUT_dim)(fit_lut3d)
# This two loss weight is also lowered to align the magnitude of loss.
loss_tv *= 0.1
loss_mn *= 0.1
loss = loss_lut_transform_image + loss_lut_regularize + loss_lut_value + loss_tv + loss_mn
acc_loss += loss.item()
# Backward.
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
# Print Log. (With est time.)
post_fix = {'acc_loss': acc_loss / (batch_idx + 1), 'loss': loss.item(),
'loss_lut_transform_image': loss_lut_transform_image.item(),
'loss_lut_regularize': loss_lut_regularize.item() if loss_lut_regularize != 0 else 0,
'loss_lut_value': loss_lut_value.item(), 'loss_tv': loss_tv.item(),
'loss_mn': loss_mn.item(), 'lr': scheduler.get_last_lr()[0]}
tqdm_bar.set_postfix(post_fix)
tqdm_bar.update(1)
logging.info(post_fix)
# Remained Time.
# tqdm_bar.set_postfix_str(tqdm_bar.postfix + " Remained Time: {:.2f}s".format(
# tqdm_bar.format_dict['elapsed'] / (batch_idx + 1) * (args.iterations // args.batch_size - batch_idx - 1)))
if args.wandb:
# Wandb Log.
wandb.log({'acc_loss': acc_loss / (batch_idx + 1), "loss": loss.item()})
wandb.log({"loss_lut_transform_image": loss_lut_transform_image.item()})
wandb.log({"loss_lut_regularize": loss_lut_regularize.item() if loss_lut_regularize != 0 else 0})
wandb.log({"loss_lut_value": loss_lut_value.item()})
wandb.log({"loss_tv": loss_tv.item()})
wandb.log({"loss_mn": loss_mn.item()})
wandb.log({"lr": scheduler.get_last_lr()[0]})
if batch_idx % 50 == 0:
# Save model.
if acc_loss / (batch_idx + 1) < loss_best:
loss_best = acc_loss / (batch_idx + 1)
torch.save(model.state_dict(), os.path.join(args.save_path_lut, 'best.pth'))
else:
torch.save(model.state_dict(), os.path.join(args.save_path_lut, 'last.pth'))
if args.save_visualization:
os.makedirs(os.path.join(args.save_path_lut, 'visualization'), exist_ok=True)
global_img = cv2.cvtColor(global_img[0].div(2.).add(0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0),
cv2.COLOR_RGB2BGR)
crop_aug = cv2.cvtColor(crop_aug[0].div(2.).add(0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0),
cv2.COLOR_RGB2BGR)
crop_mask = crop_mask[0].clamp(0, 1).cpu().numpy().transpose(1, 2, 0)
crop_gt = cv2.cvtColor(crop_gt[0].div(2.).add(0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0),
cv2.COLOR_RGB2BGR)
lut_transform_image = cv2.cvtColor(
lut_transform_image[0].detach().div(2.).add(0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0),
cv2.COLOR_RGB2BGR)
lut_transform_image = lut_transform_image * crop_mask + crop_gt * (1 - crop_mask)
crop_aug_reverse = cv2.cvtColor(
crop_aug_reverse[0].detach().div(2.).add(0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0),
cv2.COLOR_RGB2BGR)
crop_aug_reverse = crop_aug_reverse * crop_mask + crop_gt * (1 - crop_mask)
cv2.imwrite(os.path.join(args.save_path_lut, 'visualization', '{}_global_img.png'.format(batch_idx)),
global_img * 255)
cv2.imwrite(os.path.join(args.save_path_lut, 'visualization', '{}_crop_aug.png'.format(batch_idx)),
crop_aug * 255)
cv2.imwrite(os.path.join(args.save_path_lut, 'visualization', '{}_crop_mask.png'.format(batch_idx)),
crop_mask * 255)
cv2.imwrite(os.path.join(args.save_path_lut, 'visualization', '{}_crop_gt.png'.format(batch_idx)),
crop_gt * 255)
cv2.imwrite(
os.path.join(args.save_path_lut, 'visualization', '{}_lut_transform_image.png'.format(batch_idx)),
lut_transform_image * 255)
cv2.imwrite(
os.path.join(args.save_path_lut, 'visualization', '{}_crop_aug_reverse.png'.format(batch_idx)),
crop_aug_reverse * 255)
if args.wandb:
# Wandb Visualize.
wandb.log({"global_img": wandb.Image(global_img)})
wandb.log({"crop_aug": wandb.Image(crop_aug)})
wandb.log({"crop_mask": wandb.Image(crop_mask)})
wandb.log({"crop_gt": wandb.Image(crop_gt)})
wandb.log({"lut_transform_image": wandb.Image(lut_transform_image)})
wandb.log({"crop_aug_reverse": wandb.Image(crop_aug_reverse)})
print("Finish training LUT model.")
logging.info("Finish training LUT model.")